A Novel Approach to Diabetic Retinopathy Detection Using Hybrid Deep Learning Models and GAN

Abstract

Diabetic retinopathy (DR) stands as a leading cause of vision impairment globally, necessitating early and precise detection to prevent severe visual outcomes. This study presents an innovative approach to DR detection by leveraging a hybrid deep learning model that combines Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) networks, enhanced by Generative Adversarial Networks (GAN). The hybrid model capitalizes on CNN’s proficiency in feature extraction from retinal images and LSTM’s capability to understand temporal dependencies within the data, thereby improving the accuracy of DR detection. The proposed method involves the integration of a GAN framework to address the challenges of imbalanced datasets, a common issue in medical imaging. By generating high-quality synthetic retinal images, the GAN enhances the training set, thereby allowing the hybrid model to learn more effectively from limited and diverse data. This study evaluates the performance of the proposed method using a comprehensive dataset, demonstrating a significant improvement in detection accuracy. Experimental results reveal that the hybrid CNN-LSTM model, when augmented with GAN-generated data, achieves an impressive accuracy rate of 98%, markedly surpassing traditional deep learning approaches. This high accuracy underscores the potential of the hybrid model in clinical applications, offering a reliable tool for early DR detection. The findings suggest that the combination of advanced deep learning techniques and GANs not only mitigates data scarcity but also enhances model performance, providing a robust solution for diabetic retinopathy screening and diagnosis.